Transferable Implicit Solvation via Contrastive Learning of Graph Neural Networks

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Implicit solvent models are crucial for efficient and realistic biomolecular simulations.
• Developing accurate and transferable coarse-grained (CG) force fields is challenging due to limitations in parameterization and analytical expressions for potentials of mean force (PMF).

Purpose of the Study:

• To propose a machine learning-based approach using graph neural networks (GNNs) to overcome challenges in developing implicit solvent models and CG force fields.
• To derive a transferable GNN implicit solvent model from atomistic simulations.

Main Methods:

• Utilized graph neural networks (GNNs) to represent solvation free energy and potential contrasting for parameter optimization.
• Trained the GNN model on 600,000 atomistic configurations of six proteins from explicit solvent simulations.
• Evaluated the model's accuracy and transferability against state-of-the-art implicit solvent models.

Main Results:

• The GNN-based implicit solvent model achieved significantly higher accuracy in solvation free energy estimations compared to existing models.
• The model successfully reproduced configurational distributions from explicit solvent simulations.
• Demonstrated reasonable transferability of the GNN model to systems outside the training dataset.

Conclusions:

• The proposed machine learning approach effectively addresses challenges in deriving implicit solvent models and CG force fields.
• The GNN model offers a promising path towards systematically improvable and transferable models for biomolecular simulations.
• This work provides valuable insights for bottom-up development of coarse-grained models.